A NNALES DE L ’I. H. P., SECTION C
E MIL I VANOV H OROZOV
I LIYA D IMOV I LIEV
Perturbations of quadratic hamiltonian systems with symmetry
Annales de l’I. H. P., section C, tome 13, n
o1 (1996), p. 17-56
<http://www.numdam.org/item?id=AIHPC_1996__13_1_17_0>
© Gauthier-Villars, 1996, tous droits réservés.
L’accès aux archives de la revue « Annales de l’I. H. P., section C » (http://www.elsevier.com/locate/anihpc) implique l’accord avec les condi- tions générales d’utilisation (http://www.numdam.org/conditions). Toute uti- lisation commerciale ou impression systématique est constitutive d’une infraction pénale. Toute copie ou impression de ce fichier doit conte- nir la présente mention de copyright.
Article numérisé dans le cadre du programme Numérisation de documents anciens mathématiques
http://www.numdam.org/
Perturbations of quadratic
Hamiltonian systems with symmetry 1
Emil Ivanov HOROZOV
Iliya
Dimov ILIEVUniversity of Sofia, Faculty of Mathematics and Informatics, 5, J. Bourchier blvd., 1126 Sofia, Bulgaria.
Institute of Mathematics, Bulgarian Academy of Sciences, P.O. Box 373, 1090 Sofia, Bulgaria.
Vol. 13, n° 1, 1996, p. 17-56 Analyse non linéaire
ABSTRACT. - It is
proved
thatarbitrary quadratic perturbations
ofquadratic
Hamiltoniansystems
in theplane possessing
centralsymmetry
can
produce
at most two limitcycles.
Key words: Bifurcation, limit cycle, quadratic vector field, Abelian integrals.
RESUME. - Nous prouvons que les
perturbations quadratiques
arbitraires dessystemes
hamiltoniensquadratiques
etsymetrie
sur laplane peut produire
auplus
deuxcycles
limites.0. INTRODUCTION
In this paper we prove the
following
result:1 Research partially supported by Grants MM - 402/94 and MM - 410/94 of the NSF of
Bulgaria.
Annales de l’lnstitut Henri Poincaré - Analyse non linéaire - 0294-1449 Vol. 13/96/01/$ 4.00/@ Gauthier-Villars
THEOREM 1. - Let H be a
generic
cubic Hamiltonian with a central symmetry :H(-x, -~) _ -H(x, y).
Then anyquadratic perturbation of
the
corresponding
Hamiltonian systemhas at most two limit
cycles for E
smallenough.
Our interest in
systems (0.1)
was motivatedby
theproject
to obtain anexplicit
andsharp
boundc(2)
for the number of limitcycles
inquadratic
vector fields which are close to Hamiltonian ones or at least to obtain an
explicit sharp
bound for the so called Hilbert-Arnoldproblem [2]
for n = 2(see [9], [12], [15], [16], [27]
for results in thisdirection).
The Hilbert- Arnoldproblem
is formulated as follows: find an upper boundZ(n)
forthe number of isolated zeros of the
integral
where
b ( h)
is the oval(compact
connectedcomponent)
of thealgebraic
curve
H(x, y)
= handdeg H
= n +1, deg f ,
g = n(or
moregenerally,
find an upper bound
Z(n, m) provided max (deg f , deg g)
=m).
Integrals
of thistype
in studies of limitcycles
were introducedby Pontrjagin [23].
The first nontrivial case was consideredby Bogdanov [6].
The existence of an upper bound
Z(n)
was establishedby
A. Varchenko[25]
and
Khovanskii [20].
Butmaking
this boundexplicit
seems to be a difficultproblem. Quite recently,
Yu.Il’ yashenko
and S. Yakovenko[19] proved
adouble
exponential
estimate forZ(n, m) provided H
is a fixed Hamiltonian from some dense set ofgeneric
Hamiltonians:Z(n, m) 22~~m~ . Although
this bound is far from the
expected (polynomial)
one, up to now this is the best known result.More
generally,
one can considerperturbations
ofintegrable systems
instead of Hamiltonian ones. In thequadratic
case, an essentialpart
of theproblem
also consists(see [28])
incounting
the isolated zeros of certain(not necessarily Abelian) integrals. Zol~dek
hasproved [28]
that inquadratic perturbations
of Lotka-Volterrasystems
with a center thecorresponding integrals
will have at most two zeros.Among
theremaining
three classes ofquadratic systems
with a center, most of results have been obtained for the Hamiltonian case(see
e.g.[17]).
Almostnothing
is known about the other twoclasses,
where at least three limitcycles
can appear.Annales de l’Institut Henri Poincaré - Analyse non linéaire
In a recent paper
[ 16]
we found that forgeneric
cubic Hamiltonians with three saddles and one centre the exact value for bothc(2)
andZ(2)
is two.For the two
remaining
basic classes of cubic Hamiltonians(namely
thosewith one saddle and one centre and
respectively
with both two saddles andcentres)
theproblem
is still open. Ourconjecture
is that the exact bound for allgeneric
cases is two(the
same bound wasconjectured
in[28]).
Thesymmetric
Hamiltonians form a co-dimension one subset within the class of Hamiltonianshaving
two saddles and two centres. For many reasons it seems to us that thesymmetric
case is the oneallowing
to be treatedcomparatively
easier at least as far as thecomputations
are concerned.In the statement of Theorem 1 as well as in the above comments we use the notion of
"generic"
Hamiltonian. There are several definitions of this notion even forquadratic
Hamiltonian vector fields.Throughout
thispaper we say the cubic H is
generic
if thecorresponding
Hamiltonianvector field dH = 0 has a centre and does not
belong simultaneously
to any of the other
integrable
classes ofquadratic
vector fields(as
listedin
[28]).
The latter condition has asimple geometrical interpretation:
non-generic
areexactly
those Hamiltonians for which the level curves in suitable coordinates have an axis ofsymmetry.
On its handevidently
the centralsymmetry
of Hgeometrically
causes dH = 0 to becentrosymmetric
with
respect
to a suitablepoint (xo,
which we without loss ofgenerality
have chosen to coincide with theorigin.
Thegeneral
caseH(xo - x, y) = -.H(~ -
is dealt with a mere translation.One can
easily
check that after eventual linearchange
of the variableseach cubic Hamiltonian with a central
symmetry
andhaving
a centre canbe written in the form
Then
"generic"
in this casemeans ~
0. Selected level curves of H for/1 =
0, _~c
E(0,1)
and /1 = 1 are drawn inFig.
1.System (0.1 )
has attracted the interest of several other authors[4], [9], [21].
Inparticular
Bamon[4]
found a value of /1(~c
=(4 3 -1) / (4 3 + 1)
=0,227....) and
a suitableperturbation
for which around one of the foci asaddle
loop
and at least one limitcycle
can coexist. Drachman et al.[9]
proved
thatfor
close to 1 and for aspecific perturbation ( f , g)
thesystem (0.1)
has two limitcycles
around one of the foci and has no limitcycle
around the other one.(In [4], [9]
a different coordinatesystem
wasused).
Our paper covers allgeneric symmetric
Hamiltonians and all smallquadratic perturbations showing
that thesharp
estimate of the total number of limitcycles
in(0.1)
is two. Moreover from ouranalysis
it follows thatVol.
all numbers 2 and
positions
of the limitcycles permitted by
thegeneral theory
ofquadratic systems [7]
can be realized. It is also easy to describe all bifurcations when theparameters
of( f , g)
vary. Some of these results(for special perturbations)
were announced in[21] ] but apparently
withoutproofs.
In this paper we do not consider the
degenerate
case = 0 whichrequires
ananalysis
up to a fourth order in c. This case as well as the other(nonsymmetric) degenerate
cases are anobject
of another research(in progress).
Perhaps
it is worthnoticing
thatsystem (0.1)
falls into classIIIa~o
of theChinese classification
[26].
Moreprecisely
this class consists of thesystems
When
b = l = m = b = 0, n ~
0 thesystem (0.4)
is Hamiltonian andcentrosymmetric
withrespect
to thepoint ( - 2a , 2n ) .
Then asimple
corollary
from Theorem 1 is thefollowing
one:COROLLARY 1. -
In I ~
0 and smallb, l,
m;b,
system(0.4)
has no more than two limit
cycles.
The
techniques
we use toget
the results of thepresent
paper is a combination of thetechniques
from[16]
with some ideas from[11].
Inparticular
weextensively
use the notion and theproperties
of the centroidcurve.
(A
centroid curve is formedby
the mass centres of a continuousfamily
of ovals within the level curves ofH.)
There are two centroidcurves in the considered case,
corresponding
to twoperiodic
annuli. In ourstudy
theelementary
factapplies
that each intersectionpoint
of the zerodivergence
line in(0.1)
with a centroid curvecorresponds
to a zero ofI ( h)
and hence to a limit
cycle. Using
the mutualposition
of the two centroidcurves, Theorem 1 is
easily
seen to be a consequence of the fact that in thegeneric
case each of the centroid curves has a non-zero curvature.A
convexity argument
isquite
natural in situations where theexpected
number of
cycles
is two.Incidentally,
aconvexity
idea has been usedby Il’ yashenko [18]
in his paper about Abelianintegrals arising
insymmetric perturbations
of Hamiltoniansystems
withsymmetry
of order two.A more detailed sketch of our construction which we believe will
help
. the reader to overcome easier the rather technical
proof
isgiven
in thenext section.
Annales de l’lnstitut Henri Poincaré - Analyse non linéaire
1. PRELIMINARIES AND OUTLINE
OF
THE PROOFWe start with the definition of the centroid curve for any cubic Hamiltonian H
having
a centre. It is easy to see that theintegral I(h,)
from
(0.2)
can be written in the formwhere Int
(h)
is theregion
inside theoval 6(h)
and = -fx-gy.
Define the
integrals
Then the centroid
point [10]
ofregion
Int(h)
has coordinates(~(h,), ~(ja))
where ~
= =Y/M.
DEFINITION 1.1. - Let
(hl, h2~
be the maximal interval of existence of the oval~i(h) (including
theseparatrix cycle
and the centreinside).
Thenthe curve
is called
[16]
the centroid curve of thefamily
of ovals.The
importance
of theconcept
of the centroid curve lies in the fact that itsgeometry
contains thecomplete
information about thepossible
zeros ofeach
integral
=aX(h)
+ +(i.e.
for all values c~,,~,
~y and allperturbations f, g), although
the definition of Ldepends only
onH. We mention also the
following
obvious facts:a)
L is affineinvariant; b)
the
endpoints
of L are the centroid Z of the area boundedby
theseparatrix cycle
and the centre Clying inside; c)
fornon-generic Hamiltonians,
Lis a line
segment; d) M ( h ) gives
the area of and T = M’ isthe
period
function. It has beenrecently proved [8]
that T is monotoneand hence
M~~ ~
0.Specifying
Definition 1.1 to the case ofsymmetric
Hamiltonians(0.3)
wesee that there are two continuous families of ovals
8 (h)
definedrespectively
in the intervals
~-h~,
and where 0hs h~
aregiven by
the formulas
Vol.
We shall denote
by Li
andby L2
thecorresponding
centroid curves. Wepoint
out that in view of thesymmetry
of H the curveL2
isjust Li
rotatedon an
angle
In thesequel
we use L to denote any ofLl, L2.
Furtherdenote
by Ci, C2
the centres andby Sl , 82
the saddles of the Hamilto- nian(0.3).
For later use wecompute
theircoordinates, obtaining
and
respectively
=(xs, ~S) _ (-~c, -~~), C2
=-G’~, ~‘2
=The
following
theorem sums up severalproperties
of the centroid curveof
generic
Hamiltonians obtained in[16]
and which we need in thepresent
paper.THEOREM 1.1. -
Suppose
that H isgeneric.
Then(i)
Near itsendpoints
the curve L isregular,
i.e.(h)
+r~’2 (h) ~
0.(ii)
Near theendpoints of
L the curvature » is not zero and has the samesign.
It is not hard to see
[ 16]
that for agiven perturbation f, g
in(0.1)
with( -I- ~,~ j ~ 0,
the number of limitcycles
in(0.1)
for small c isequal
to thenumber of the intersection
points (counting
themultiplicities)
of L with theline £ : ax +
-~- ~y
=0, provided
L isregular.
Moreover as c --~ 0 theselimit
cycles
tend to the ovals8(h)
whose centroidpoints
lie on £. Henceour
goal
would be achieved if we show thatLi, L2
areregular
and that any line intersects their union L =L1
UL2
in at most twopoints.
In section 5we prove the
regularity
of the centroid curves in the case of anarbitrary generic
Hamiltonian with two centres,using
Picard-Lefschetztheory [3].
The
proof
is similar to that in[16]
andexploits
ideas from[12]
and[22].
Note that the
regularity
of L as well as thenonvanishing
of the curvatureare affine invariant conditions. Most of our efforts will be concentrated to show that an
arbitrary
line £ intersectsLi
in no more than twopoints.
Therefore the
following
theorem is the bulk of our construction:THEOREM 1.2. - The curvature
of
the centroid curveLl
at eachpoint
is non-zero.
Our
plan
to show thenonvanishing
of the curvature is to start with aHamiltonian H for which we
already
know that. Such one can be obtained forexample
from the standardelliptic
HamiltonianHo == y2 -~- x3 - ~
via smallperturbation
in the direction of thesymmetric
Hamiltonians. This is done in section 6. From Theorem 1.1 we know that in thegeneric
casethe curvature near the
endpoints
of the centroid curve isalways
non-zero.Annales de l ’lnstitut Henri Poincaré - Analyse non linéaire
This
implies
thatvarying
theparameters
in H if the curvatureacquires
azero this would be at least a double zero. It will
correspond
to a zero ofmultiplicity
at least four ofI ( h) provided f, g
areproperly
chosen([16], Corollary 2.1).
Asimple
butimportant
fact is that the second derivative of I and the first derivative of theperiod
function Tsatisfy
Picard-Fuchssystem
of order two, and hence the ratio w =I" /T’
satisfies a Riccatiequation.
Thisobservation,
used for the first time in[ 11 ],
is based on the fact that the residua of the form w= - ,~ dy + 9 dx
are linear in handtherefore the second covariant derivative of the Gauss-Manin connection of w has no residua. In section 2 of the paper we use these ideas to derive Riccati
equation
for w. Most of thepresent
paper is devoted to aprecise qualitative study
of the Riccatiequation
satisfiedby
w. Inparticular
weexplore
theproperties
of the zero isocline of theequation.
On its hand tostudy
the zero isocline we include it in afamily
of curvesforming
the levelcurves of a suitable Hamiltonian. We
repeat
the sameprocedure
this timeconsidering
the horizontal and also the vertical isocline of the Hamiltoniansystem
to obtain itsphase portrait.
From theproperties
of the zero isoclineof the Riccati
equation
we find thepossible
behaviour of w. Theupshot
is that w =
I" /T’ (and
henceI ~~ )
has no double zeros. All thisanalysis
is contained in section 4.
We have to
emphasize
that in section 4 we do notstudy w
for all valuesof the
parameters
a,j3,
~y. We are interestedonly
in values a,corresponding
to lines ftangent
toLi.
In this case w issubject
toquite
restrictiveinequalities,
which we obtain in section 3. Theproof
of Theorem 1 is
completed
in section 7by examining
the mutualposition
of the two centroid curves.
In
theory
the sameplan
could beapplied
to all another cases(including
the one studied in
[16]). Unfortunately
thecomputations
ingeneral
becomemuch more
complicated.
Inspite
of that we believe that theplan
can beperformed, eventually
with some modifications. In thisrespect
thepresent
paper,apart
of its ownimportance,
can serve as a model for thegeneral
case.2. PICARD-FUCHS SYSTEM AND RICCATI
EQUATION
In this section we derive several differential
equations
satisfiedby
Abelianintegrals.
Our finalgoal
will be the Riccatiequation
satisfiedby I"
For this we need to derive Picard-Fuchssystem
for theintegrals Io, I_ 1, 1-2 ,
defined below.
Given h E
(-he, -hs),
we consider the level curve H =xy2
+3 x3 - x
=h,
whoseequation
can be written also in the formPut z = xy +
M/2
and define theintegrals
We have
and
similarly Y(h)
=M(h) _ - I-1.
LEMMA 2.1. - The
integrals Io, I-1, ~-2 satisfy
thefollowing
systemof equations
Proof. - Express
the left and theright
hand-side of theidentity
in terms of the functions
Ik, k
=0, ~ 1, ...
Thisgives
a connectionbetween these functions:
Similarly using
Annales de l’lnstitut Henri Poincaré - Analyse non linéaire
we obtain
Using (2.3)
with k = -1 and(2.4)
with k = 0 weget
the thirdequation
of
(2.1).
In the same mannerapplying (2.2)
and(2.3)
with k = 0 and(2.4)
with k == -1 we obtain the secondequation
of(2.1).
At theend,
ifwe eliminate
Ii
and1-3
from(2.2), (2.3), (2.4)
with k= -2,
we obtain the firstequation.
COROLLARY 2.2. - The
integrals X, Y,
Msatisfy
the systemNow we are in a
position
to derive the Riccatiequation
satisfiedby w(h)
= Put v = w - ~. Then we can consider thatv =
I~~/M~~
where inI(h)
wehave ~
= 0. Differentiate once the firsttwo
equations
in(2.5)
and twice the third one toget
From these
equations
and from I = aX eliminate first Y and then X.After
solving
withrespect
to7
andM
we obtainVol.
Here for shortness we have denoted
A i /
where
From
(2.6)
we derive the Riccatiequation
Here B =
B1
+B2. Returning
to the variable w = v-~ -y
andwriting (2.8)
as a
system
wefinally get
LEMMA 2.3. - The
following
Picard-Fuchs system issatisfied:
this is our basic
system
which we willstudy
in the next two sections. Inparticular
of agreat importance
for ouranalysis
will be theproperties
of thezero isocline
Fo of (2.9), given by
thepoints (h, w )
on thealgebraic
curve(2.10)
We will often consider
Fo
as thegraph
of the two-valued functionwo (h)
determined
by P ( h, wo ( h) )
- 0("+"
isalways assigned
to the upperbranch).
For later use denote
by Do
the discriminant of thequadratic equation (2.10):
Annales de l’lnstitut Henri Poincaré - Analyse non linéaire
3. BASIC
INEQUALITIES
In this section we derive several
inequalities concerning
theparameters
in(2.9)
and the values ofwo
at the critical values ofH, provided
the lineP : ax +
{3y + "y
= 0 and the centroid curveLi satisfy
some conditionslisted below. These
inequalities
are crucial in ouranalysis
of(2.9).
Through
this and the next section we suppose thefollowing:
(U) (i) ~
is atangent
to the centroid curveLi
at an internalpoint
~~~jL~, rl(h,~~~
(ii)
the curvature ofLi
does notchange
thesign (but
mayvanish);
(iii) running L1,
£ rotates on anangle
less than ~r.We start with certain restrictions for the coefficients of the line £
imposed by (U).
LEMMA 3.1. - The
coefficients
a,,~, ~y (multiplied if needed by -1) satisfy
the
inequalities
Proof. -
We aregoing
to prove that theangular
coefficient of £ satisfies theinequalities
If this was done and
assuming
thatj3
> 0 weimmediately get
the lasttwo
inequalities
in(3.1).
Moreover theloop through Si
isplaced
in thehalf-plane x
> 0 and it iseasily
verified that if(~,
and(.x, y2)
arepoints
on
8 (h),
h E(-h~. -ja5),
then ~/i + y2 0. ThereforeLi
is situated in the fourthquadrant
and since 03B1 0j3
theequality
++ -y
= 0 ispossible only if 03B3
> 0.In order to prove
(3.2)
we need another normal formof the Hamiltonian. In these coordinates
Li corresponds
to values h E[0, ~].
Having
in mind(U),
theproof
consists offinding
theequations
of thetangents
at theendpoints
ofLi. According
toCorollary
3.4 from[16]
the
equation
of thetangent .~~
ofLi
at the centreCl(O,O)
-Ax(see Fig. 2).
Thetangent .~s
at the otherendpoint
goesthrough
the saddleS 1
=(1, 0)
andthrough
the centroid of theloop
areaZi [16],
but thecoordinates of
Zi
aregiven by extremely long
formulas. Because of thiswe use another line instead of
.~s .
Thesaddle-loop
is determinedby
theequation
H =~.
The line x= 2
divides theregion
inside the saddle-loop
into twoparts Ac
ands containing respectively
the centre and thesaddle.
Obviously
the reflection ofAs through
the line x= 2
is contained inAc.
This meansthat ~ ( 6 ) ~.
Thenusing (U)
and results from[16]
yields
thatZi
=(~ ( 6 ) , ~ ( 6 ) )
is located inside thetriangle
with vertices( 0, 0 ) , ( 2 , 0 ) , ( 2 , - 2 ) .
Let m be the linesymmetric
to.~~
withrespect
to the line x= 2 ,
m has anequation
y= ~ ( x - 1).
Then theangular
coefficient k of an
arbitrary tangent £
at an interiorpoint
ofLi
satisfiesAnnales de l ’lnstitut Henri Poincaré - Analyse non linéaire
Fig. 2. _
the
inequalities -A 1~
A. In order to return to the coordinatesystem
in which the normal form is(0.3)
wechange
the variables(x, y)
-(xl,
where
In this coordinate
system
theequations
of.~~
and m becomerespectively (we
omit thesubscript 1)
Hence the
angular
coefficient k of anytangent
line satisfies~ ~
1/~c.
Vol. 13, n°
As a direct consequence of
(2.7)
and(3.1 )
we obtainCOROLLARY 3.2. - The constants
A; B, C, D,
E in(2.9)
arepositive.
LEMMA 3.3. -
(i)
The valuesofw/ ( h)
at the endsof the
intervalhave the
following signs:
(ii)
The derivativessatisfy
Proof. - (i) Using
theexplicit
formulas forA, A,
B etc. we express the discriminantDo
from(2.11)
in the formAlso we have
This
gives
after obvious calculationsAnnales de l ’lnstitut Henri Poincaré - Analyse non linéaire
Further,
with thehelp
of(2.7)
and the formulas for~~
from section 1we obtain via
straightforward computations
thatwo ( - ~~ ~
=~x~ -f-,~3~~ -~-~y.
In the same way we
compute
At the endwhere
From
(3.4)
it is clear that thepoint ~~)
islying
on thetangent .~~.
We will show further that
~~
xs.Really
xs = -~~ and~~
+5~ 1 - I~2.
But x~ + ~/c5h~
isequivalent
toThe last
inequality easily
follows fromh~ 9 (2
+3~) ~~ ( ~
+~c) .
The
inequality ~~
shows that~e)
is. above anytangent
.~ toLi
whichgives
that = ++ ~
> 0. The other twoinequalities
in(i)
follow from the fact that both the centre~~)
and thesaddle
(xs,
are below anytangent
.~.(ii) Evaluating
the derivatives ofwo
at A = 0 weget
Write
Do
in the formThis
gives
Calculation the values of A’
at - hs and -he yields:
Finally
we substitute h= - h~
in the above formulas and via directcomputations
findRepeating
the abovecomputation
for h= - hs
we obtainthus
proving
the lemma.4. THE ZERO ISOCLINE OF THE RICCATI
EQUATION
In this section we list a number of
properties
of the zero isoclineFo
of(2.9) supposing (U)
hold. Inparticular
decisive for our purposes will be the number and the mutualpositions
of the minima and maxima ofwo ( h) .
We start with the
following
lemma.LEMMA 4.1. -
(i)
The curveFo is centrosymmetric
with respect to thepoint (0,~).
..(ii) Fo
has no compact components.Proof -
Assertion(i)
is obvious. To prove(ii)
we observe that the discriminantDo given by (2.11)
can have no more than two zeros for h 0.Together
with the behaviour near h = 0(see
the nextlemma),
thiseliminates the existence of
compact components.
Annales de l’lnstitut Henri Poincaré - Analyse non linéaire
Next we
study
certainasymptotic properties
of h 0.LEMMA 4.2. -
(i)
Thefunction wo (h)
has thefollowing asymptotic expansions
(ii)
Thefunction wo (h)
has thefollowing asymptotic expansions :
Proof. -
Directcomputations.
Our further
plan
is to consider the entirefamily
of level curveswhich is tantamount to the
phase portrait
of the Hamiltoniansystem
This is because the discriminant
Do
is eitherpositive
or canchange
thesign depending
on theparameter
values in(2.11 ). Consequently Fo
can bifurcateand all
possible
situations should be occurred in thefamily (Fp :
p EI~~.
Again
forsimplicity
of the notation we can consider theportrait for 7
=0,
the
general
casebeing
a mere translation w -~ w + ~y.LEMMA 4.3. -
System (4.1)
hasexactly
onesingular point So
in thehalf-plane h
0. Thepoint So
is a saddle.Proof -
From theequations Pw
=Ph
= 0 eliminate w andput h2
= zto obtain the
equation
Obviously
thisequation
has at least onepositive
root zo. Theanalysis
ofthe coefficient alternations and Descartes’ criterion rules out the existence
Vol.
of other
positive
roots. As the Poincare index of the field in(4.1)
cannotexceed
0, using
thesymmetry yields
thatSo
is a saddlepoint.
Denote
by .ho == -~o
the abscissa of the saddleSo .
Our furtheranalysis
will concentrate
mainly
onstudy
of two of the isoclines of(4.1), namely
the horizontal
© ~
and the vertical one V= ~ Pw
=0 ~ .
The
properties
we need are easy to obtain and are summed up in thefollowing
lemma:LEMMA 4.4. -
(i)
The curve V is concave in thehalf-plane
h 0.(iii)
Theasymptotics
holdProof -
Directcomputation.
’The horizontal isocline ?-~ is more
complicated
forstudying.
Moreover it bifurcates asparameters
vary. To describe itsproperties
we denoteby
~~~~~h),
two roots of theequation Ph (h, w)
=0,
h 0.LEMMA 4.5. -
(i)
Let~~ -
3AE > 0. Then consistsof
two separatecurves
given by
thegraphs of wH.(h),
h 0. Theirasymptotics
are the
following
Each
of
thefunctions w~. ( h)
hasexactly
one maximum and no minima.Let
.~2 -
3AE 0. Then the twographs of w~
meet at afinite point
thus
fornfng geometrically
one smooth curve H. For the twobranches have the same
asymptotics
as in(4.2).
Thefunction
hasalways
one maximum. Thefunction w~ (h~
has either one minimum and onemaximum or no extrema, the latter case
including
also thepossibility of
acollision
of
the maximum and minimum.Proof. -
Theasymptotics
arecomputed straightforward
from the formulasAnnales de l’Institut Henri Poincaré - Analyse non linéaire
When B2 - 3AE ~
0 both exist for all h 0. WhenB2 - 3AE
0 the discriminant D has asimple
zero at anegative
value h =h,
i.e.=
wH(h)
and both are defined for h E(-oo, h].
At the endcomputing
the derivative of we obtain:Putting h2
= z thisyields
a cubicequation
for all extrema:In the case
B2 -
3AE > 0 theasymptotics (4.2)
say that each of has at least one maximum. As(4.3)
has at most twopositive solutions,
this proves
(i).
In the case
B2 -
3AE 0 theequation (4.3)
has either one or threepositive
solutions.Using
theasymptotics (4.2)
and the behaviour near h show that has at least one maximum while either has noextrema or has one maximum and one minimum. To obtain more
precise information,
we observe that the derivative ofw~
isnegative
for zjB/2
and the derivative of
~7~
ispositive
for z >B / 2.
IfF ( z )
denotes thepolynomial
in(4.3),
we haveF ( B ) _ - 3AE 2 0, F’ ( B ) _ - 8B E
0.Hence
(4.3)
hasexactly
one zerogreater
thanB/2.
Thisyields
thatwjj
hasalways exactly
one extremum(maximum).
Lemma 4.5 isproved.
We will denote the two branches of H
by H+
and ?-~- .Our next
objective
is to locate the saddleSo
withrespect
to the extrema of Before that we need more refinestudy
ofH,
which we doagain by considering
the wholefamily
of curvesPh
= const and their horizontal and vertical isoclinesHH = ~ Ph h
=4 ~
andHv
=0 ~ .
Denote alsoby
?-~C~~..r
the two branches ofHH (see Fig. 3).
As above we consider these curves asgraphs
of thecorresponding
functionswH(h)
andwHv(h), h
0.LEMMA 4.6. -
(i)
The curves andHv
are concave. Eachof H±H
hasa maximum at =
6
and has a maximum athHV = B .
(ii)
The curvesHH
andHv
have no commonpoint
andHv
is situatedbetween the two branches
of HH.
Proof -
Directcomputation.
LEMMA 4.7. -
(i)
The curves V andHv
intersect at theunique point of
maximum
of
V .Fig. 3.
(ii)
The curves V and have no commonpoint,
V andH-H
intersect ata
unique point S*
with an abscissah* satisfying
(iii)
The saddleSo
is on the lower branchof
H. Its abscissaho satisfies
Proof - (i)
is obvious.(ii)
To prove the existence anduniqueness
ofS*
we argue as above in
establishing
the existence of the saddleSo.
For theproof
of theinequality hHH h*
we first take theasymptotic
values ofthe functions
wHH(h)
andwir (h)
when ~ --~-0, namely
Annales de l’Institut Henri Poincaré - Analyse non linéaire
This shows that
wHH(h)
>wv(h)
near h = -0. On the other hand wecan see that which in details reads
or
(3B2
+36C)2 4B2(4B2
+6ABD). Using
theexpressions
forA, B, C, D
weeasily verify
that 12CB2,
B AD. Then(3Bz
+36C)2 3684 4B2(4B2
+SABD) 4B2(4B2
+6ABD).
This proves
(4.4).
(iii)
We followessentially
the sameargument, showing
thatThis condition is
equivalent
to(B2 -~-12C)2 8B2(8B2 +
9ABD -6AE)
which follows from 12C
B2
and theeasily
verifiedinequality
AEB2.
The
inequality (4.6)
weproved implies
that V and the lower branch H- intersect at apoint So
with an abscissaho satisfying (4.5). Really
ifB2 -
3AE > 0 then(4.2)
and Lemma 4.4(iii) give
that >wv (h)
near h = -0. If
B2 -
3AE 0 then(4.6) implies
thathHv h,
where his the zero of D from Lemma 4.5
(ii).
Hence =wHv(h)
>wv(h)
because h >
hHv
>hv
andwHv(h)
>wv(h)
for h >hv,
the latter aconsequence of
(i),
Lemma 4.4(ii)
and Lemma 4.6(i).
ThushHv ho h
and
(iii)
isproved.
Another
property
of H and V we need is thefollowing
LEMMA 4.8. -
Suppose
that thefunction wH (h)
has a maximum athH.
Then
hH ho (The
saddleSo
is to theright of the
maximumof w~(h)).
Proof. -
The extrema of H- lie on~CH
and~C~
itself is a concave curve with a maximum at apoint
with an abscissahHH.
Sincethis
yields
that the maximum of ~C-(if
itexists)
has an abscissahH hHH
and moreoverwH (h)
forh
Suppose
thatho hH.
Then = and= > because >
wv(h)
for h >ho.
Therefore the intersection
point S*
of V andH h
has an abscissah*
withho h* hH hHH -
a contradiction with(4.4).
This proves the lemma.In order to make the facts
proved
in Lemmas 4.3-4.8 moreapplicable,
below we list all needed
points
of H and V with their coordinates and characteristics:’
- the saddle
point, So
= H nV,
the
point
of maximum of~+, Sv(hv , wv (hv)) -
thepoint
of maximumof V,
the
point
of maximum of~-,
the
point
of minimum of~-C-,
~’(h, wH (h) ) -
thepoint
whereH+
meets ~-~C’ .(The
last threepoints
do not existnecessarily.)
Now we are able to describe in details the
picture
of the minima and maxima ofwo ( h ) .
LEMMA 4.9. -
(i)
LetDo
>0 for
any h. Then thefunction
h 0has
exactly
one maximum and one minimum.(ii)
LetDo
0for h
E(h-, h+),
h-h+
0. Thena) for
h E
( - oc , h - ~
thefunction wo ( h )
hasexactly
one maximum and thefunction wo {h)
hasexactly
one maximum and oneminimum; b) for
h E
~h+, 0)
thefunction wo (h)
hasexactly
one maximum and one minimum.Proof -
Theseparatrices
H and V divide the halfplane h
0 into four open domains ifB2 -
3AE 0 and into five open domains otherwise. Letus denote
by ~l~
andVi
theparts
of ?~C- and Vrespectively
which are tothe left of the saddle
So
andby H2
andV2 -
the restpart
of H and V.Analytically (for example)
we have= ~ ( h, wH ( h) ) :
hho}
andsimilarly
for the other arcs. In the case whereB2 -
3AE >0, ~-C2
itself consists of twoseparate
curves, one of them isH+
and the other is thepart
of to theright
of the saddle. We will write sometimes and~-C2
inorder to
signify them,
thusH2
=H+2
n?nC2
in this case. Denoteby Oij
thedomain in h 0 which
boundary
isHi
UVj, i, j
=1, 2. (More precisely
in the case
B2 -
3AE > 003A922
consists of twoseparate
subdomains:5222
with
boundary H-2
UV2
and03A9+22
withboundary U {h
=0};
in the caseB2 -
3AE 0 theboundary
off222
is~C2
UV2 U {h
=0~ .)
Therefore
depending
on thesign
ofB2 -
theposition
of the saddlepoint
and theavailability
of extrema on ~-C ~ there are four cases I - IV(see Fig. 4, a,b,c,d):
I
B2 -
3AE0,
~C- has no extrema,II
B2 -
3AE0, So
is betweenS~
andS,
III
B2 -
3AE0, So
is betweenSH
and,~~,
IV
B2 -
3.4E > 0.Annales de l’Institut Henri Poincaré - Analyse non linéaire